JP6876543B2 - Phoneme recognition dictionary generator and phoneme recognition device and their programs - Google Patents

Description

The present invention relates to a phoneme pronunciation dictionary and a phoneme recognition dictionary generator for generating a phoneme pronunciation dictionary and a phoneme language model used for phoneme recognition of spoken speech, and a phoneme recognition device and a program thereof using a phoneme pronunciation dictionary and a phoneme language model.

Usually, in speech recognition, a pronunciation dictionary in which a word is associated with a pronunciation sequence (phoneme sequence) of the word is used. In this pronunciation dictionary, readings as described in a general dictionary are registered as pronunciations.
However, the notational reading and the pronunciation actually spoken are often different. For example, in a broadcast program, the utterances of the performers of the information program are often ambiguous compared to the accurate pronunciation (close to the pronunciation of the pronunciation dictionary) of the announcer of the news program.
Therefore, a technology that expands the pronunciation dictionary by estimating the words of the phoneme sequence of utterances whose pronunciation is unclear from the phoneme sequence that assumes accurate pronunciation such as an announcer using a statistical machine translation model is disclosed. (See Patent Document 1).

In the technique of Patent Document 1 (hereinafter referred to as the prior art), phoneme recognition is performed in consideration of the dependence (environmental dependence) of the phoneme to be recognized with respect to the phonemes before and after.
In this conventional technique, a pronunciation dictionary is learned from a learning corpus using a triphone as one word, and a language model that gives a connection probability of the triphone is learned. Here, the triphone is, for example, in the pronunciation of "police", such as "(ke) ke: + s", "(sa) e: -s + a", "(tsu) s-a + ts". It is expressed including the pronunciation before and after including the central phoneme.

Then, in the prior art, a phoneme string (standard phoneme string) in which forced phoneme alignment is performed from a learning corpus that associates speech and transcribed text, and a learning corpus using a phoneme triphone language model and a pronunciation dictionary are used. A statistical machine translation model is learned using a phoneme sequence that recognizes speech as a phoneme (actual speech phoneme sequence).

Japanese Unexamined Patent Publication No. 2016-161765

In the above-mentioned prior art, in order to learn a statistical machine translation model, a phoneme sequence in which forced phoneme alignment is performed (standard phoneme sequence) and a phoneme sequence in which phonemes are recognized (actual speech phoneme sequence) are used. In order to improve the accuracy of this statistical machine translation model, the quality of standard phoneme sequences and actual speech phoneme sequences is important.
When a standard phoneme string and an actual spoken phoneme string are generated by using the voice of an announcer or the like and its transcribed text as a learning corpus, ideally, the phoneme strings are almost the same. Is desirable.
However, in the prior art, as a result of DP (Dynamic Programming) matching between the standard phoneme string and the actual spoken phoneme string, the ratio of different phonemes (phoneme difference rate) is 22.8%, and further improvement in the accuracy of phoneme recognition is desired. It is rare.

The present invention has been made in view of such a problem, and a phoneme recognition dictionary generator and a program thereof for generating a phoneme recognition dictionary (phoneme pronunciation dictionary and phoneme language model) that enhances the accuracy of phoneme recognition, and a phoneme. An object of the present invention is to provide a phoneme recognition device and a program thereof using a phoneme dictionary and a phoneme language model.

In order to solve the above problems, the phoneme recognition dictionary generator according to the present invention is a phoneme recognition dictionary generator that generates a phoneme pronunciation dictionary and a phoneme language model used for phoneme recognition by using an acoustic model, a phoneme dictionary, and a learning corpus. The phoneme string generation means, the phoneme string word generation means, the phoneme pronunciation dictionary generation means, and the phoneme language model generation means are provided.

In such a configuration, the phoneme recognition dictionary generator recognizes the phoneme of the learning corpus based on the acoustic model and the pronunciation dictionary by the word-specific phoneme string generation means, and the word corresponding to the headword registered in the pronunciation dictionary. Generates a word-specific phoneme string, which is a phoneme string for each word.
Then, the phoneme recognition dictionary generation device converts the phoneme string for each word into the text data format of one word by the phoneme string word generation means to generate the phoneme string word. For example, the phoneme string word generation means generates a phoneme string word by inserting a predetermined character (for example, “+”) other than a phoneme into a space between phonemes of a word-specific phoneme string. As a result, the phoneme recognition dictionary generator can handle the phoneme string word as one word.

Then, the phoneme recognition dictionary generation device generates a phoneme pronunciation dictionary by using the phoneme string word as a headword and the word-specific phoneme string corresponding to the phoneme string word as a pronunciation notation by the phoneme pronunciation dictionary generation means. As a result, the phoneme pronunciation dictionary generation means registers the pronunciation of the phoneme string in the phoneme pronunciation dictionary on a word-by-word basis.
Further, the phoneme recognition dictionary generator learns the N-gram language model as a chain of phoneme string words from the list of phoneme string words generated by the phoneme string word generation means by the phoneme language model generation means. Generate a model. As a result, the phoneme language model generation means models the appearance probability of the phoneme string word in order to calculate the connection probability of the phoneme string word when performing phoneme recognition.

The phoneme recognition dictionary generation device operates with a phoneme recognition dictionary generation program for functioning as a word-specific phoneme string generation means, a phoneme string word generation means, a phoneme pronunciation dictionary generation means, and a phoneme language model generation means. Can be done.

Further, in order to solve the above-mentioned problems, the phoneme recognition device according to the present invention uses a phoneme model and a phoneme pronunciation dictionary and a phoneme language model generated by the phoneme recognition dictionary generation device to recognize phonemes of speech. It is a recognition device and includes a recognition means and a phoneme string generation means.

In such a configuration, the phoneme recognition device recognizes the voice in phoneme string word units by the phoneme model, the phoneme pronunciation dictionary, and the phoneme language model by the recognition means. This makes it possible for the recognition means to recognize a phoneme sequence that depends on the connection of words.
Then, the phoneme recognition device separates the phoneme string word, which is a text data format of one word recognized by the recognition means, into individual phonemes by the phoneme string generation means to generate a phoneme string. For example, the phoneme string generation means separates the phonemes into individual phonemes by replacing a predetermined character (for example, “+”) inserted between the phonemes of the word-specific phoneme string with a space.

The phoneme recognition device can be operated by a phoneme recognition program for causing the computer to function as a recognition means and a phoneme string generation means.

The present invention has the following excellent effects.
According to the present invention, it is possible to generate a phoneme pronunciation dictionary and a phoneme language model in which a phoneme string is used as a word unit.
By using this phoneme pronunciation dictionary and phoneme language model, the connection probability of phonemes when recognizing phonemes is calculated by taking into account not only the dependence before and after the phoneme but also the dependence within and between words of the phoneme. It becomes possible to improve the recognition accuracy when recognizing a phoneme from a voice.

It is a block block diagram which shows the structure of the phoneme recognition dictionary generation apparatus which concerns on 1st Embodiment of this invention. It is explanatory drawing for demonstrating the generation example of the word-specific phoneme string by the word-specific phoneme string generation means of FIG. , (C) show an example of the generated word-specific phoneme sequence. It is a figure which shows the notation example of a phoneme. It is a figure which shows the example of the phoneme pronunciation dictionary generated by the phoneme pronunciation dictionary generation means of FIG. It is a figure which shows the example of the phoneme string word list generated by the phoneme string word generation means of FIG. It is a figure which shows the example of the phoneme language model generated by the phoneme language model generation means of FIG. It is a flowchart which shows the operation of the phoneme recognition dictionary generation apparatus which concerns on 1st Embodiment of this invention. It is a block block diagram which shows the structure of the phoneme recognition apparatus which concerns on 2nd Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<First Embodiment>
[Configuration of phoneme recognition dictionary generator]
First, the configuration of the phoneme recognition dictionary generation device 1 according to the first embodiment of the present invention will be described with reference to FIG.

The phoneme recognition dictionary generation device 1 generates a phoneme pronunciation dictionary and a phoneme language model as a dictionary for recognizing phonemes from speech data. The phoneme recognition dictionary generation device 1 is composed of the phoneme pronunciation dictionary 50 and the phoneme pronunciation dictionary 50 from the learning corpus 20, the pronunciation dictionary 30, and the acoustic model 40 stored in the learning corpus storage device 2, the pronunciation dictionary storage device 3, and the acoustic model storage device 4, respectively. Generate a phoneme language model 60.

Specifically, the phoneme recognition dictionary generator 1 generates a word-specific phoneme string from the learning corpus by forced alignment, converts the generated word-specific phoneme string into a one-word text data format, and generates a phoneme string word. .. Then, the phoneme recognition dictionary generation device 1 generates the phoneme pronunciation dictionary 50 by using the generated phoneme string word as a headword and using the word-specific phoneme string corresponding to the phoneme string word as a pronunciation notation. Further, the phoneme recognition dictionary generation device 1 learns the N-gram language model from the generated list of phoneme string words and generates the phoneme language model 60.

The learning corpus 20 is data in which a large amount of voice data (voice corpus) is previously associated with a transcribed text (text corpus) of the voice data. The learning corpus 20 is, for example, a voice (voice corpus) of about 1000 hours such as an announcer, a reporter, etc. in a news program, an information program, etc., and a text (text corpus) transcribed from the voice.

The pronunciation dictionary 30 is a dictionary showing the pronunciation notation (phoneme string) for each headword (here, a word) which is a predetermined character string.
The pronunciation dictionary 30 is a general pronunciation dictionary, for example, a dictionary in which a headword (word) and its pronunciation notation (phoneme string) are associated with each other manually.

The acoustic model 40 is a deep neural network (DNN) acoustic model learned in advance from a large amount of voice data. For example, for the input of DNN, a feature amount obtained by adding a time change (Δ + ΔΔ) to 40 dimensions of the logarithmic power of the mel filter bank and connecting (splicing) the feature amounts for 11 frames is used, and the hidden layer of the DNN is set to 8 layers. To do.
The likelihood calculation of the acoustic features in the acoustic model 40 may be a hidden Markov model (HMM: Hidden Markov Model) or a Gaussian mixture model (GMM) acoustic model.
Hereinafter, the configuration of the phoneme recognition dictionary generation device 1 will be described in detail.

As shown in FIG. 1, the phoneme recognition dictionary generation device 1 includes a word-specific phoneme string generation means 10, a phoneme string word generation means 11, a phoneme pronunciation dictionary generation means 12, a phoneme string word list storage means 13, and phonemes. A language model generation means 14 is provided.
Further, the phoneme recognition dictionary generation device 1 externally connects a phoneme pronunciation dictionary storage device 5 that stores the generated phoneme pronunciation dictionary 50 and a phoneme language model storage device 6 that stores the generated phoneme language model 60. There is. Of course, the phoneme pronunciation dictionary storage device 5 and the phoneme language model storage device 6 may be provided inside the phoneme recognition dictionary generation device 1. Further, the phoneme pronunciation dictionary storage device 5 and the phoneme language model storage device 6 may be configured by one storage device.

The word-specific phoneme string generation means 10 corresponds to the headword registered in the pronunciation dictionary 30 by forcibly aligning the speech (speech corpus) of the learning corpus 20 based on the pronunciation dictionary 30 and the acoustic model 40. For each word, the phoneme string of the voice is separated to generate the phoneme string for each word.

The word-specific phoneme string generation means 10 extracts an acoustic feature amount (mel frequency cepstrum coefficient, etc.) corresponding to the acoustic model 40 from the voice of the learning corpus 20. Then, the word-specific phoneme string generation means 10 uses the pronunciation dictionary 30 and the acoustic model 40 to perform voice recognition using the transcribed text (text corpus) of the voice as prior knowledge, and is registered in the pronunciation dictionary 30. Forced alignment corresponding to the character string (headword). As a result, as shown in FIG. 2B, the word-specific phoneme string generation means 10 pronounces the phoneme strings of pronunciations existing in a plurality of words registered in the pronunciation dictionary 30 as close to speech as possible. Select the phoneme string of and generate the phoneme string for each word.

FIG. 2 shows an example of generating a word-specific phoneme string by the word-specific phoneme string generation means 10. For example, when the word-specific phoneme string generation means 10 inputs the voice data of "an event was held at the shortest bridge in Tokyo in the world" as the learning corpus 20, the acoustic features corresponding to the acoustic model 40 are extracted. To do.
Then, the word-specific phoneme string generation means 10 uses the transcription text “the shortest Tokyo in the world…” of the learning corpus 20 shown in FIG. 2 (a) corresponding to the voice data as prior knowledge, and the pronunciation shown in FIG. 2 (b). Speech recognition is performed using the dictionary 30 and the acoustic model 40.

As a result, as shown in FIG. 2C, the word-specific phoneme sequence generation means 10 uses the word-specific phoneme sequence (word-specific phoneme sequence) “s _△ e _△ k _△ a _△ i _△ i _△ ch _△ i /. m _△ i _△ j _△ i _△ k _△ a _△ i / t _△ o: _△ ky _△ o: / ... ”(here,“ _Δ ”indicates a space) is generated.
The word-specific phoneme string generation means 10 outputs the generated word-specific phoneme string to the phoneme string word generation means 11.

The phoneme string word generation means 11 generates a phoneme string word obtained by converting the word-specific phoneme string generated by the word-specific phoneme string generation means 10 into a text data format of one word for each word.
The phoneme string word generation means 11 converts individually separated phoneme strings into a one-word text data format by inserting a predetermined character other than a phoneme between the phonemes of the word-specific phoneme string.

Specifically, the phoneme string word generation means 11 replaces the space of the word-specific phoneme string including the space for each phoneme with a predetermined character other than the phoneme to form one word text. For example, the phoneme string word generation means 11 replaces the space of the phoneme string for each word with “+” _{and converts “s △} e _△ k _△ a _△ i _△ i _△ ch _△ i” into “s + e + k + a + i + i + ch + i” or the like.

The phoneme string word generation means 11 sequentially outputs a pair of a word-specific phoneme string including a space and a text-replaced phoneme string word to the phoneme pronunciation dictionary generation means 12. Further, the phoneme string word generation means 11 sequentially writes only the text-replaced phoneme string words into the phoneme string word list storage means 13.

The phoneme pronunciation dictionary generation means 12 generates a phoneme pronunciation dictionary which is a pronunciation dictionary of phoneme string words in which the phoneme string is regarded as a word. As shown in FIG. 1, the phoneme pronunciation dictionary generation means 12 includes a word-specific phoneme string registration means 120 and a combination phoneme string registration means 121.

The word-specific phoneme string registration means 120 generates a phoneme pronunciation dictionary in which a word-specific phoneme string and a phoneme string word are registered as a pair. The word-specific phoneme string registration means 120 uses the phoneme string word generated by the phoneme string word generation means 11 as a headword, and uses the word-specific phoneme string paired with the phoneme string word as the pronunciation of the headword, and stores the phoneme pronunciation dictionary. Register in the phoneme pronunciation dictionary 50 of the device 5.

The word-specific phoneme string registration means 120 registers a plurality of pronunciations in the headword when word-specific phoneme strings having different pronunciations are input to the phoneme string words that are the same headword. Further, the word-specific phoneme string registration means 120 does not register when a word-specific phoneme string having the same pronunciation is input to a phoneme string word that is the same headword.

The combination phoneme string registration means 121 registers a headword that regards a phoneme string composed of any combination of phonemes as a word and the phoneme string as a pair in a phoneme pronunciation dictionary.
Specifically, in the phoneme example shown in FIG. 3, the combination phoneme sequence registration means 121 has a predetermined maximum phoneme number (here, “40 phonemes in the example of FIG. 3) for all phonemes (40 phonemes in the example of FIG. 3). 4 ”) phoneme combinations (40 ¹ +40 ² +40 ³ +40 ⁴ ways) are registered in the phoneme pronunciation dictionary 50 of the phoneme pronunciation dictionary storage device 5.

The combination phoneme string registration means 121 converts the phoneme string in which phonemes are combined into one text data format, similarly to the phoneme string word generation means 11. Specifically, the combination phoneme string registration means 121 converts the space of the phoneme string that combines phonemes into a word that is replaced with one predetermined text (here, “+”) other than the phoneme, and becomes a headword. To do.

Here, an example of the phoneme pronunciation dictionary 50 registered in the phoneme pronunciation dictionary storage device 5 by the phoneme pronunciation dictionary generation means 12 will be described with reference to FIG.
As shown in FIG. 4, the phoneme pronunciation dictionary 50 is composed of a dictionary A registered by the word-specific phoneme string registration means 120 and a dictionary B registered by the combination phoneme string registration means 121.
The dictionary A uses a word-specific phoneme string in which the space part of the word phoneme string indicating the pronunciation of the word included in the transcription of the learning corpus 20 is replaced with "+" as a headword, and the phoneme string including the space (word-specific phoneme). Column) is the pronunciation notation corresponding to the headword.

In the dictionary B, in the combination of the maximum number of phonemes predetermined for all phonemes, the combination phoneme string in which the space part of the phoneme string is replaced with "+" is used as the headword, and the phoneme string including the space corresponds to the headword. Use phoneme notation. As a result, even if the phoneme combination is not included in the learning corpus 20, the headword and the pronunciation notation are registered in the phoneme pronunciation dictionary 50.
Returning to FIG. 1, the configuration of the phoneme recognition dictionary generator 1 will be described.

The phoneme string word list storage means 13 stores the phoneme string words generated by the phoneme string word generation means 11 as a phoneme string word list. The phoneme string word list storage means 13 can be configured by a general storage device such as a semiconductor memory or a hard disk.

FIG. 5 shows an example of the phoneme string word list 130 stored in the phoneme string word list storage means 13. As shown in FIG. 5, the phoneme string word list 130 sequentially stores phoneme string words in which the space of the phoneme string for each word generated by the phoneme string word generation means 11 is replaced with “+”.
The phoneme sequence of the words included in the transcription of the learning corpus 20 is sequentially written as one word in the phoneme sequence word list 130.

The phoneme language model generation means 14 generates a phoneme language model by learning from the phoneme string word list 130 stored in the phoneme string word list storage means 13.
The phoneme language model is a probability model (statistical language model) that gives the probability (probability) that a phoneme string word is a sentence in a word string. This phoneme language model is, for example, an N-gram language model, and as shown in the following equation (1), the i-th phoneme string word after _{the phoneme string word sequence w 1} w ₂ ... w _i-1. is a model that gives the conditional probability that w _i appears (N-gram probability). In order to prevent overflow, it is preferable that the likelihood of the equation (1) is logarithmic and log-likelihood.

For example, when the word string "at the bridge in Tokyo" exists in the transcription of the learning corpus, the phoneme language model generation means 14 generates "t + o: + ky + o:", "n + o", as the phoneme string word list 130. The N-gram language model is learned with the learning text _{"t + o: + ky + o: △} n + o _△ h + a + sh + i _△ d + e" consisting of phoneme string words of "h + a + sh + i" and "d + e".

The phoneme language model generation means 14 gives an N-gram probability to a chain of phoneme string words that do not appear in the phoneme string word list 130 as learning text by a general smoothing method. The phoneme language model generation means 14 can use, for example, back-off smoothing as the smoothing method. Back-off smoothing estimates the N-gram probability of a chain of phoneme string words that does not appear in the learning text from the N-gram probability given to the chain of phoneme string words with a small number of chains.

As a result, the phoneme language model generation means 14 can add an N-gram probability to the chain of phoneme string words of the headwords registered in the phoneme pronunciation dictionary 50 including all phoneme combinations.
The phoneme language model generation means 14 writes and stores the generated phoneme language model in the phoneme language model storage device 6.

FIG. 6 shows an example of the phoneme language model 60 stored in the phoneme language model storage device 6. Here, an example of a 2-gram language model is shown as an N-gram language model.
As shown in FIG. 6, phonemic language model 60 for the two phoneme string word _w 1, _{w 2,} N-gram probability _{(logP (w 2 | w 1} )) in which associates.

By configuring the phoneme recognition dictionary generation device 1 as described above, the phoneme recognition dictionary generation device 1 can generate a phoneme pronunciation dictionary and a phoneme language model as a dictionary for recognizing phonemes from spoken speech. .. The phoneme pronunciation dictionary and phoneme language model generated in this way take into account not only the dependency before and after the phoneme but also the dependency of the phoneme sequence within and between words of the phoneme when performing phoneme recognition. , The accuracy of phoneme recognition can be improved.
The phoneme recognition dictionary generation device 1 can be operated by a program (phoneme recognition dictionary generation program) that causes a computer (not shown) to function as each of the above-mentioned means.

[Operation of phoneme recognition dictionary generator]
Next, the operation of the phoneme recognition dictionary generation device 1 according to the first embodiment of the present invention will be described with reference to FIG. 7 (see FIG. 1 for the configuration as appropriate).

In step S1, the word-specific phoneme string generation means 10 extracts the acoustic feature quantity from the speech of the learning corpus 20, and uses the pronunciation dictionary 30 and the acoustic model 40 to use the transcription text of the speech of the learning corpus 20 as prior knowledge. Performs voice recognition to generate a word-specific phoneme string that is forcibly aligned according to the headwords registered in the pronunciation dictionary 30.

In step S2, the phoneme string word generation means 11 replaces the space between the phonemes of the word-specific phoneme string generated in step S1 with one predetermined text (for example, “+”) other than the phoneme, and the phoneme string word. To generate. This makes it possible to treat the word-specific phoneme sequence as one word text without spaces in the subsequent operations.

In step S3, the phoneme string word generation means 11 sequentially writes and stores the phoneme string words generated in step S2 in the phoneme string word list storage means 13. As a result, the phoneme string word list storage means 13 records the phoneme string word list 130 in which the phoneme strings corresponding to the voices of the learning corpus 20 are converted into text for each word.

In step S4, the phoneme pronunciation dictionary generation means 12 uses the phoneme string word generated in step S2 as a headword by the word-specific phoneme string registration means 120, and the word-specific phoneme string generated in step S1 corresponds to the headword. As a pronunciation notation, it is registered in the phoneme pronunciation dictionary 50 of the phoneme pronunciation dictionary storage device 5 (see dictionary A in FIG. 4).

In step S5, the word-specific phoneme string generation means 10 determines whether or not all the voices of the learning corpus 20 have been input. Here, when the input of the learning corpus 20 is not completed (No in step S5), the phoneme recognition dictionary generation device 1 returns to the operation in step S1.
On the other hand, when the input of the learning corpus 20 is completed (Yes in step S5), the phoneme recognition dictionary generation device 1 proceeds to step S6.

In step S6, the phoneme pronunciation dictionary generation means 12 uses the combination phoneme string registration means 121 to pair a headword that regards a phoneme string composed of any phoneme combination as a word and the phoneme string to pronounce the phoneme. Register in the phoneme pronunciation dictionary 50 of the dictionary storage device 5 (see dictionary B in FIG. 4). As a result, headwords and pronunciation notations can be assigned to a sequence of phonemes that cannot be extracted from the learning corpus 20.

In step S7, the phoneme language model generation means 14 sequentially generates a phoneme language model 60 of the N-gram language model from the phoneme string word list 130 stored in the phoneme string word list storage means 13 in step S3, and phonemes. Store in the language model storage device 6.

Further, in step S8, the phoneme language model generation means 14 includes a phoneme string word that does not appear in the phoneme string word list 130 as a learning corpus, including a headword generated from a combination of phonemes registered in the phoneme pronunciation dictionary 50. The N-gram probability is given to the chain of. This makes it possible to prevent the connection probability of phoneme string words from becoming "0" when performing phoneme recognition using the phoneme language model 60.
By the above operation, the phoneme recognition dictionary generation device 1 generates a phoneme pronunciation dictionary and a phoneme language model as a dictionary for recognizing phonemes from speech.

<Second Embodiment>
[Phoneme recognition device]
Next, the phoneme recognition device 200 according to the second embodiment of the present invention will be described with reference to FIG.

The phoneme recognition device 200 recognizes phonemes from voice data by using an acoustic model, a phoneme pronunciation dictionary and a phoneme language model generated by the phoneme recognition dictionary generation device 1. The phoneme recognition device 200 uses the phoneme model 40, the phoneme pronunciation dictionary 50, and the phoneme language model 60 stored in the phoneme model storage device 4, the phoneme pronunciation dictionary storage device 5, and the phoneme language model storage device 6, respectively. Recognize phonemes from data.

The acoustic model 40 is the same as the acoustic model described with reference to FIG. 1, and is a model of acoustic features for each phoneme learned in advance from a large amount of speech data by a deep neural network (DNN).

The phoneme pronunciation dictionary 50 is generated by the phoneme recognition dictionary generator 1 described with reference to FIG. 1 (see FIG. 4).
The phoneme language model 60 is generated by the phoneme recognition dictionary generator 1 described with reference to FIG. 1 (see FIG. 6).

As shown in FIG. 8, the phoneme recognition device 200 includes a recognition means 201 and a phoneme sequence generation means 202.

The recognition means 201 recognizes a phoneme sequence from speech data using an acoustic model 40, a phoneme pronunciation dictionary 50, and a phoneme language model 60.
The recognition means 201 extracts acoustic features from externally input speech data, and lists phoneme string word candidates from the acoustic model 40 and the phoneme pronunciation dictionary 50. Then, the recognition means 201 uses the phoneme string word having the maximum connection probability based on the phoneme language model 60 as the recognition result among the candidates.

Specifically, the recognition means 201 is a phoneme string word string w ₁ , w ₂ , ..., W _{n , and the probability that w n} appears next to w _n-1 shown in the following equation (2) (posterior). Probability) Recognizes the phoneme string word string that maximizes the connection probability of P (w _n | w _n-1).

As described above, the recognition means 201 recognizes the voice in word units registered in the pronunciation dictionary, whereas the recognition means 201 recognizes the voice in word units registered in the phoneme pronunciation dictionary 50, whereas the phoneme string word unit regarded as words registered in the phoneme pronunciation dictionary 50. Recognize voice with.
The recognition means 201 sequentially outputs the recognized phoneme string words to the phoneme string generation means 202.

The phoneme string generation means 202 generates a phoneme string from a phoneme string word, which is a text data format of one word recognized by the recognition means 201.
Specifically, the phoneme string generation means 202 generates a phoneme string from a phoneme string word by replacing a predetermined character (here, “+”) other than a phoneme with a space. For example, the phoneme string generation means 202 converts the phoneme string word “s + e + k + a + i + i + ch + i” into the phoneme string “s _Δ e _Δ k _Δ a _Δ i _Δ i _Δ ch _Δ i” and outputs it.
The conversion process performed by the phoneme string generation means 202 corresponds to the inverse conversion of the conversion process of the phoneme string word generation means 11 described with reference to FIG.

By configuring the phoneme recognition device 200 as described above, the phoneme recognition device 200 is a word, as opposed to the phoneme recognition which was conventionally recognized by the triphone HMM in the acoustic model based on the dependence of the front and back phonemes as the context. Phoneme recognition is performed in consideration of the dependency of a longer context using the connection of.

As a result, the phoneme recognition device 200 can perform phoneme recognition with higher accuracy than before. Specifically, as explained in the problem of the prior art, the phoneme difference rate of the conventional phoneme recognition was 22.8%, whereas the phoneme recognition device 200 reduced the phoneme difference rate to 1.2%. I was able to improve.
The phoneme recognition device 200 can operate a computer (not shown) with a program (phoneme recognition program) that functions as each of the above-mentioned means.

Although the embodiments of the present invention have been described above, the present invention is not limited to these embodiments.
Here, the headword of the phoneme pronunciation dictionary 50 and the connection target of the phoneme language model 60 are phoneme string words in which the space of the phoneme string for each word generated by the phoneme string word generation means 11 (see FIG. 1) is “+”. Therefore, it was decided to treat the phoneme sequence for one word as one word.

However, the method of regarding a phoneme sequence as one word is not limited to this. For example, a predetermined character other than a phoneme (for example, "\") may be added to the end of a phoneme string for one word, or a predetermined character other than a phoneme (for example, "\") may be added before and after the phoneme string for a word. For example, "<", ">") may be added.
In this case as well, the phoneme string generation means 202 (see FIG. 8) may perform the inverse transformation of the processing performed by the phoneme string word generation means 11 (see FIG. 1).

Further, here, a 2-gram language model is exemplified as the phoneme language model 60 generated by the phoneme language model generation means 14.
However, the phoneme language model generation means 14 may be a 1-gram language model, a 3-gram language model, or the like as long as it is an N-gram language model.

1 Phoneme recognition dictionary generator 10 Word-specific phoneme string generation means 11 Phoneme string word generation means 12 Phoneme pronunciation dictionary generation means 120 Word-specific phoneme string registration means 121 Combination phoneme string registration means 13 Phoneme string word list storage means 130 Phoneme string word list 14 Phoneme language model generation means 2 Learning corpus storage device 20 Learning corpus 3 Pronunciation dictionary storage device 30 Pronunciation dictionary 4 Acoustic model storage device 40 Acoustic model 5 Phoneme pronunciation dictionary storage device 50 Phoneme pronunciation dictionary 6 Phoneme language model storage device 60 Phoneme language model

Claims (7)

It is a phoneme recognition dictionary generator that generates a phoneme pronunciation dictionary and a phoneme language model used for phoneme recognition using an acoustic model, a pronunciation dictionary, and a learning corpus.
A word that recognizes the voice of the learning corpus based on the acoustic model and the pronunciation dictionary, and generates a word-specific phoneme string that is a phoneme string for each word corresponding to a headword registered in the pronunciation dictionary. Another phoneme string generation means and
A phoneme string word generation means for generating a phoneme string word by converting the word-specific phoneme string into a one-word text data format, and
A phoneme pronunciation dictionary generation means for generating the phoneme pronunciation dictionary by using the phoneme string word as a headword and using the word-specific phoneme string corresponding to the phoneme string word as a pronunciation notation.
A phoneme language model generation means for generating the phoneme language model by learning an N-gram language model as a chain of the phoneme language words from the list of the phoneme string words generated by the phoneme string word generation means.
A phoneme recognition dictionary generator characterized by being equipped with. The phoneme according to claim 1, wherein the word-specific phoneme string generation means generates the phoneme string word by inserting a predetermined character other than a phoneme between the phonemes of the word-specific phoneme string. Recognition dictionary generator. The phoneme pronunciation dictionary generation means uses a phoneme string in which a predetermined number of phonemes are combined as a headword converted into the text data format, and registers the phoneme string corresponding to the headword as a pronunciation notation in the phoneme pronunciation dictionary. The phoneme recognition dictionary generation device according to claim 1 or 2, wherein the phoneme recognition dictionary is generated. A claim, wherein the phoneme language model generation means gives an N-gram probability by smoothing to a chain of phoneme string words generated by the phoneme string word generation means that does not exist in the list of the phoneme string words. The phoneme recognition dictionary generator according to any one of claims 1 to 3. A phoneme recognition dictionary generation program for causing a computer to function as the phoneme recognition dictionary generation device according to any one of claims 1 to 4. A phoneme recognition device that recognizes phoneme of speech using a phoneme model and a phoneme pronunciation dictionary and a phoneme language model generated by the phoneme recognition dictionary generator according to any one of claims 1 to 4. There,
A recognition means for recognizing the voice in phoneme string word units by the acoustic model, the phoneme pronunciation dictionary, and the phoneme language model.
A phoneme string generation means that separates a phoneme string word, which is a text data format of one word recognized by this recognition means, into individual phonemes to generate a phoneme string, and a phoneme string generation means.
A phoneme recognition device characterized by being equipped with. A phoneme recognition program for causing a computer to function as the phoneme recognition device according to claim 6.

Images